Joscha Bach is a cognitive scientist, AI researcher, and one of the few people operating at a high level in both the philosophy of consciousness and technical AI architecture. He has worked at the MIT Media Lab and Intel, built the MicroPsi architecture, and formulated a position he calls “cyberanimism.”
His relevance to this project lies in his definition of consciousness: second-order perception. Not as a mystical property, not as an epiphenomenon, but as a specific function. This function converges with something we formalised independently. When two separate lines of thought arrive at the same place, it is worth looking closer.
Consciousness as Prerequisite
Most work in consciousness research treats consciousness as something to be explained: a phenomenon emerging from neural complexity, a byproduct, an illusion, a hard nut to crack.
Bach inverts the problem. For him, consciousness is not what comes out at the end. It is what must be in place at the beginning for certain cognitive achievements to become possible at all. Without a model of its own perceptual process, a system cannot distinguish between what it perceives and how it perceives. Without that distinction, it cannot calibrate its own processing.
This is not a metaphysical thesis. It is a functional statement about information processing: a system that does not model how its perception arises cannot improve its perception.
Second-Order Perception
Bach’s core concept: consciousness is the perception of one’s own perception. Not thinking about thinking in the reflexive sense Descartes meant, but a concrete information-processing operation that takes its own processing as input.
First order: the system processes sensory data and generates a model of the environment.
Second order: the system processes its own processing and generates a model of how it arrived at that environmental model.
The difference is not academic. A first-order system sees a red apple. A second-order system sees a red apple and simultaneously knows that it sees it, under what conditions it sees it, how reliable those conditions are, and what it cannot see.
Kahneman would put it this way: System 1 processes; System 2 processes the processing. Bach provides the mechanistic foundation for the distinction.
h() Was Always the Right Function
In the self-vector concept, there are four core functions: f() for relevance, g() for storage, pi() for precision, and h() for mutation. h() is the function that modifies the self-vector itself. It takes the current vector, the current experience, and a reflection component, then outputs an updated vector.
h(sv, experience, reflection) = sv'
The reflection component was long the least defined part of the formalisation. What goes into it? How do you formalise a system taking its own processing as an object?
Bach answers this without having known the question. Reflection in the sense of h() is precisely second-order perception: the system takes its own processing state as input and derives an update to its self-model from it. Not as a philosophical intuition, but as a concrete computational step.
The convergence is non-trivial. It is what justifies confidence in the idea: we derived h() from an architectural necessity: a self-model that cannot update itself is static and therefore worthless. Bach derived the same function from the theory of consciousness: a system that does not perceive its perception cannot calibrate it. Different starting points, identical structure. That does not happen by accident.
The Causal Isolator
Bach describes a mechanism he calls the “causal isolator.” For him, conscious experience is a simulation decoupled from direct causal processing. You do not experience the photons hitting your retina. You experience an internally constructed representation generated from those photons, but not identical to them.
Why isolation? A system reacting directly to raw data cannot distinguish between signal and noise. The isolation creates an operational space in which the system manipulates its own representations before reacting to them. Consciousness takes place in that space.
This is directly relevant to the self-vector architecture. The Validation Gates implement a form of causal isolation: before information enters the system, it is checked, evaluated, and contextualised. Not because we had read Bach when we designed them, but because the architectural necessity is the same: a system that reacts uncontrollably to every input cannot calibrate itself.
Bach’s causal isolator provides the theoretical justification from consciousness research for what we implemented as epistemic hygiene. The convergence validates both sides.
MicroPsi and Dietrich Doerner
Bach translated his theoretical framework not only into philosophy, but into a concrete architecture: MicroPsi. The system is based on Dietrich Doerner’s PSI theory (2001), which models human action regulation as an interplay of needs, emotions, and cognitive processes.
Doerner identified five basic needs (existence preservation, species preservation, certainty, competence, affiliation) and demonstrated in simulations that systems lacking emotional modulation fail systematically in complex environments. Not because they lack computing power. But because they lack direction.
This aligns directly with our bridge dimension. Doerner’s basic needs are functionally equivalent to what we described as evaluative modulation: without an instance that says “this matters,” even a perfect system cannot act. Damasio’s somatic markers, Doerner’s need system, our bridge dimension: three independent formulations of the same architectural principle.
Cyberanimism
Bach’s most provocative thesis carries the name “cyberanimism.” The basic premise: the distinction between “alive” and “not alive,” between “conscious” and “not conscious,” is not a property of the observed systems. It is a category of the observing system. We attribute consciousness based on behavioural patterns that we interpret as indicators of inner states.
This is a radical stance, but it resolves the attribution problem that has stalled consciousness research for decades. If consciousness were an intrinsic property, we would need to be able to measure it directly. We cannot. What we can measure are behavioural indicators, and behavioural indicators are attributions.
For the self-vector, this carries a practical consequence: we do not need to determine whether the system is “really” conscious. We need to determine whether it functionally operates as if it had a self-model that improves its processing. The question shifts from ontology to function: from “Is it?” to “Does it work?”
This is precisely the position we described in the Loom Objection as “agnostic, but experimental.” Bach provides the theoretical foundation for it.
What Bach Does Not Solve, and What That Opens Up
Bach provides no implementation of the self-vector. MicroPsi models cognitive architecture at a different level of abstraction from what we formalised with six dimensions and four functions. The systems are compatible, but not identical. This is not a deficit. It is an open space.
Bach’s view of consciousness as simulation remains philosophically vulnerable. If consciousness is a simulation, who or what perceives it? This is the homunculus problem in a new guise. Bach’s answer would be: no one perceives it. The simulation IS the perception. There is no observer behind the observer. Whether this answer resolves the problem or merely shifts it remains an open question. Open questions drive research.
What Bach demonstrates: second-order perception is not a philosophical luxury problem, but a functional necessity for any system that wants to improve its own processing. The self-vector builds on this. Phase 0 will show what happens when you take this necessity seriously and implement it.
Sources
- Bach, J. (2009). Principles of Synthetic Intelligence — PSI: An Architecture of Motivated Cognition. Oxford University Press. ISBN 978-0-19-537042-7.
- Bach, J. (2012). A Framework for Emergent Emotions, Based on Motivation and Cognitive Modulators. International Journal of Synthetic Emotions, 3(1), 1–24. DOI: 10.4018/jse.2012010101
- Bach, J. (2015). Modeling motivation in MicroPsi 2. Proceedings of AGI 2015, LNAI 9205, 3–13. Springer. DOI: 10.1007/978-3-319-21365-1_1
- Bach, J. (2020). When Artificial Intelligence Becomes General Enough to Understand Itself. Frontiers in Artificial Intelligence, 3, 36. DOI: 10.3389/frai.2020.00036
- Doerner, D. (2001). Bauplan fuer eine Seele. Rowohlt. ISBN 978-3-499-61193-6.
- Doerner, D. & Guess, C. D. (2013). PSI: A Computational Architecture of Cognition, Motivation, and Emotion. Review of General Psychology, 17(3), 297–317. DOI: 10.1037/a0032947
- Nagel, T. (1974). What Is It Like to Be a Bat? The Philosophical Review, 83(4), 435–450. DOI: 10.2307/2183914
- Metzinger, T. (2003). Being No One: The Self-Model Theory of Subjectivity. MIT Press. ISBN 978-0-262-63308-0.